Enhancing knowledge tracing through course map and question difficulty analysis
Jianing Xia, Amanda Li, Hui Yin, Guozhe Jin · Array · 2025
Knowledge tracing (KT) has enhanced students’ learning efficiency in online systems by modeling how students learn over time based on their interactions with educational content. It has gained significant research attention recently due to its crucial importance in intelligent education. Nonetheless, current knowledge tracing (KT) methods have not adequately explored and utilized the impact of concept relations and question difficulty level on students’ knowledge state during learning. This study aims to address this gap by investigating the effect of the concept map and question difficulty on learning and proposing the Course Map and Difficulty-enhanced Knowledge Tracing (CMDKT) model to enhance students’ knowledge state assessment. Experiments on ASSIST09 and DBE-KT22 show CMDKT improves predictive performance over strong baselines (e.g., +2.47% AUC and +2.32% ACC on average across datasets), while also reducing MAE and RMSE compared to DKT and GKT. • CMDKT model integrates concept maps and question difficulty for enhanced knowledge tracing By modeling both prerequisite concept relationships and difficulty levels, CMDKT improves accuracy in predicting students’ knowledge states. • CMDKT outperforms state-of-the-art models like DKT and GKT across datasets Achieved up to 2.47% improvement in AUC and 2.32% increase in accuracy, demonstrating strong generalizability across both ASSIST09 and DBE-KT22 datasets. • CMDKT captures learning complexity via dual GRU architecture One GRU models concept dependencies (topology embedding), while the other captures temporal knowledge progression for better personalization. • CMDKT shows consistent gains as more concept prerequisites are included Performance improves steadily (e.g., ACC from 0.7845 to 0.7976) as the percentage of concept prerequisite pairs increases, validating the importance of relational learning structure.